Xiaohongshu's OSDI Paper: All-Flash ANNS System Cuts 90% Costs

机器之心 · wechat · 2026-07-22

Xiaohongshu's engine architecture team proposed HELMSMAN, a high-performance Approximate Nearest Neighbor Search (ANNS) system optimized for all-flash servers, with their paper accepted by OSDI 2026.

Core Problem: Traditional in-DRAM graph indexes (like HNSW) are prohibitively expensive, while existing DRAM-SSD hybrid systems struggle with strong SLA and high QPS due to serial I/O bottlenecks.
System Design: HELMSMAN leverages cluster-based indexing and builds a user-space storage stack using SPDK to bypass Linux kernel overhead. It introduces Leveling-Learned Search Pruning (LLSP) to adaptively predict the search range before reading, enabling batched I/O to maximize SSD bandwidth.
Build Acceleration: It utilizes GPUs to accelerate coarse-grained k-means clustering and an elastic CPU pool for fine-grained balancing, enabling 10B-scale index reconstruction within hours.
Business Impact: Using only 40 all-flash servers, it replaced a workload previously requiring 35,000 CPU Cores and 350TB of DRAM, saving over 90% in hardware costs. It achieved a 2-16x throughput improvement compared to systems like DiskANN, reaching up to 85% of pure DRAM throughput.

Original post →

Xiaohongshu's OSDI Paper: All-Flash ANNS System Cuts 90% Costs

More from Infra